Networks and the Dynamics of Firms' Export Portfolio:
Evidence for Mexico
Juan de Lucio
Raúl Mínguez
Asier Minondo
Francisco Requena
ISSN 1749-8368
SERPS no. 2014003
February 2014
Networks and the dynamics of firms' export portfolio: Evidence for Mexico
Juan de Lucio (High Council of Spanish Chambers of Commerce, Spain)
Raúl Mínguez (High Council of Spanish Chambers of Commerce, Spain)
Asier Minondo (Deusto Business School, Spain)
Francisco Requena (University of Sheffield, United Kingdom) *
Abstract
In this paper we use network-analysis tools to identify communities in the web of exporters'
destinations. Next we use our network-based community measure as predictor of additional
countries chosen by firms expanding their export destination portfolio. We defend that our
network-based community measure is superior to extended gravity measures. This superiority
stems from the fact that community is a revealed measure, is country-specific and can be
calculated at the industry level. Using data on Mexican new exporters over the period 20032009, we show that the probability of choosing a new export destination multiplies almost by
three if it belongs to the same community of any of the firm’s previous destinations. The
introduction of the network-based community variable improves the accuracy of the model up
to 20% relative to a model that only includes gravity and extended gravity variables. We also
show that industry-specific communities and general communities play similar roles in
determining the dynamics of Mexican exporters' country portfolio.
JEL Code: F1
Keywords: export market, network analysis, modularity, extended gravity, Mexico.
Acknowledgements: We thank Juanjo Gibaja, Laura Rovegno and participants at research
seminars at the Inter-American Development Bank, University of Murcia, University of
Sheffield and the World Bank, and at XVI Encuentro de Economía Aplicada in Granada, the
XIV Congress in International Economics in Palma de Mallorca and the XV ETSG
Conference in Birmingham. We also acknowledge financial support from the Spanish
Ministry of Economy and Competitiveness (MINECO ECO2010-21643 and ECO201127619, co-financed with FEDER), the Basque Government Department of Education,
Language policy and Culture and Generalitat Valenciana (project Prometeo/2011/098). We
thank the World Bank for sharing with us the database on Mexican exporters.
*Corresponding author: Francisco Requena. Department of Economics, University of
Sheffield, Edgar Allen House, 241 Glossop Road, Sheffield S10 2GW UK. Tel: + 44 114 222
3398, Fax: + 44 114 222 3458. Email: [email protected]
1
Introduction
Product and country export portfolio is much more concentrated in developing
countries than it is in developed countries (Cadot et al, 2013). In addition, survival rather than
entry into exports markets is the key to understand the growth of exports in developing
countries (Besedes and Prusa, 2011). These two facts motivate this paper: there is a need to
examine the dynamics of firms’ export destination portfolio in developing countries. If
destination path-dependence in exports exists, it is important that firms choose adequately
export destinations because the chances of survival are higher and, by extension, export
growth.
Until recently firm-level research in this area has mostly treated export status as a
binary variable: firms are either exporting or they are not. Hence, empirical studies of entry
into exporting have focused on the initial entry decision, particularly on identifying the firmspecific characteristics which set exporting firms apart from non-exporters. In this paper we
focus our enquiry on a subsequent question: given that firms have the ability to export, what
determine their choices about where to export?
Exporters do not add new destinations to their portfolio at random. In fact, firms have
a higher probability of adding destinations that are similar to their domestic market. This
similarity is governed by the so-called gravity factors, such as income, geography or culture.
Firms also have a higher probability to add new destinations that are similar to their previous
destinations. Previous literature contends that this similarity can be approached by the socalled extended-gravity variables that capture the geographical and cultural closeness between
previous and new destinations (Morales et al., 2011).
This paper is divided into two parts. In the first part we develop a new similarity
measure using the tools of complex network analysis. In particular, we identify communities
of countries sharing similar characteristics within the web of export markets of Mexico. Our
measure has several advantages relative to extended gravity indicators. First, it is a revealed
measure, and hence it captures not only all extended-gravity proxies, but also any nonmeasurable or non-observable characteristic that might also affect how similar export
destinations are. Second, the community measure is a country-specific measure. In the case of
extended gravity variables, the similarity between two export destinations is always the same,
irrespective of the location of the exporter. For example, the (extended gravity) distance
between Country A and Country B is the same for exporters from Mexico and another
Country (say C). However, it might be the case that Country A and Country B belong to the
2
same community of Mexico but not of Country C. For example, Country A and Country B
might have the same preferences for Mexican products but not for products from Country C.
The third advantage of our network-based measure is that we can identify industry-specific
communities. It might be the case that countries belong to the same community in one
industry but to a different community in another industry. For example, in the case of Mexico
and regarding tequila, Country A and Country B might form a community because both
countries have the same regulations on the maximum alcohol content. In this case, the tequila
that has been modified to meet the requirements in Country A will also be suitable for
Country B, leading these two countries to form a community. In contrast, for book sales,
Country A and Country B might belong to different communities because they speak different
languages. Extended-gravity measures cannot control for these differences because
geographical and cultural variables do not usually vary across industries.
In the second part of the paper, we use our network-based similarity measure as
determinant of the destinations-portfolio of new regular Mexican exporters over the period
2003-2009. We show that an exporter will have almost three times higher probability of
choosing a new destination if it belongs to the same community of any of its previous
destinations. The network variable keeps its strong predictive capacity even when we control
for gravity and extended-gravity variables, and improves the accuracy of the model up to
20%. We also identify industry-specific destination-communities and general destinationcommunities and find that both have a similar influence on the evolution of Mexican firms’
export-portfolio.
Our analysis is related to different strands of literature that analyze firms' exports
dynamics. Chaney (2011) proposes a model of international network formation where firms
obtain information about new potential partners from their current trading partners. The
network formation game yields equilibria where firms' export destinations are pathdependent. Albornoz et al. (2012) and Nguyen (2012) develop alternative multi-market export
models based on the idea that a firm's foreign demands are uncertain and correlated across
markets. When faced with multiple destinations to which they can export, many firms will
choose to sequentially export in order to slowly learn more about its chances for success in
untested markets. Experimentation becomes an optimal strategy leading to path-dependence
in firms' export destinations.
Our paper is also related to the concept of "the geographic spread of trade", a term
originally proposed by Evenett and Venables (2002). They showed that geographic and
linguistic proximity to an existing export-market was a consistently significant factor in
3
determining expansion into new markets for sector-level exports from developing countries,
implying a role for learning from existing export experiences. Using firm-level export data,
Morales et al. (2011) for Chile, Lawless (2011) for Ireland, and Defever et al. (2011) for
China explore the role of “extended gravity” forces and show that firms tend to choose new
export destinations that are similar (geographically, culturally or economically) to
destinations that firms are already exporting to.
Our paper is also related with the novel literature that applies network methods to
analyze international trade. Kali and Reyes (2007) map the topology of international trade and
develop new measures of economic integration based on network analysis. De Benedictis and
Tajoli (2011) apply network analysis to describe the evolution of international trade and to
study other trade-related topics. Hidalgo et al. (2007) use the probability of exporting
products in tandem to develop a measure of proximity between products, which is displayed
into a map using network analysis. These authors show that the product-map determines the
evolution of countries' productive specialization. Based on the product-map, Kali et al. (2013)
show that countries' growth prospects are enhanced if their export basket is closer, measured
using network analysis tools, to more complex products that the country does not export yet.
The remainder of the paper is organized as follows. Section 2 presents the data and
applies network analysis tools to identify communities within the web of Mexican exporters'
destinations. Section 3 introduces the empirical model to examine how connectedness
between countries affects the choice of new destinations by Mexican new exporters that
expand their export portfolio. Section 4 presents the results of the empirical analyses. Section
5 concludes.
2. Communities in the network of Mexican export markets.
In this section, we start presenting the database used to perform our empirical analysis
and then explore the web of export destinations in Mexico using tools from network analysis.
Next, we explain how to identify communities in a network and apply the identification
algorithm of communities to the entire network of Mexican exporters. Finally, we construct
industry-specific networks and implement the algorithm of identification of communities
separately to each of them and show that the number of communities and its members can
vary from industry to industry.
2.1 An exploration of the web of export destinations.
4
We use the transaction level customs data on the universe of Mexican exporters over
the period 2000-2009. The database was facilitated by the World Bank's Trade and Integration
Team (Cebeci et al, 2012). 1 The database provides the annual value of exports per firm,
destination and Harmonized System 6-digit product code.2 We use data over the period 20032009 to analyze the dynamics of exporters’ destination portfolio in the next section (our
dependent variable in the econometric exercise) and data over the period 2000-2002 to
examine the network of export markets and construct the community measure in this section
(our main explanatory variable). A detailed description of the Mexican firm-level data as well
as other data sources used in the paper can be found in Appendix 1.
We begin our empirical analysis by examining the network of Mexican firms' export
destinations using tools from network analysis. Figure 1 presents the network of Mexican
exporters' destinations in year 2002.3 Export destinations are nodes in the web and two nodes
are connected by an edge if there is at least a firm that exports to both nodes. The size of the
node is correlated with the number of firms that export to that destination and the size of the
edge (weight) is correlated with the number of firms that export to both destinations linked by
the edge. The network has 175 nodes and 9,576 edges. The density of the network is 0.63. 4
The most important destinations for Mexican exporters were the US (25,730 firms),
Guatemala (2,534 firms), Canada (1,931 firms), Costa Rica (1,855 firms) and El Salvador
(1,394 firms). The edges with the highest weights were Canada-US with 1,534 firms
exporting to both destinations, Guatemala-US with 1,368 firms, Costa Rica-US with 1,084
firms, Costa Rica-Guatemala with 923 firms and Colombia-US with 913 firms. All the nodes
are connected in the network; this means that there is no destination where all exporters to
that destination only exported to that destination. Each node has an average degree of 55; that
is, the total number of additional destinations served by firms that export to a destination is,
on average, 55. As expected, the destination with the highest degree is the US: 170 edges. It is
followed by Chile (164), Canada (160), Guatemala (159) and Colombia (158).
2.2 Identification of communities.
In the destination network, we want to identify communities of countries that have
stronger relationships among them than with the rest of destinations in the network. These
1
Data were collected by the Trade and Integration Unit of the World Bank Research Department as part of their
efforts to build the Exporter Dynamics Database (http://econ.worldbank.org/exporter-dynamics-database).
2
For example, one record of our database is the annual value of exports of “Pullovers, cardigans etc. of wool or
hair, knit” (HS-6 code 6110101) by a Mexican exporting firm (identifier 15) to Italy in 2002.
3
In the robustness section we use 2000 and 2001 data.
4
If all nodes were connected with each other density would be 1.
5
tighter relationships reveal that if a firm exports to one country in the community it will also
tend to export to other countries within the community. One widely used procedure to identify
communities within a network is the maximization of a modularity function (Newman, 2006),
which is expressed as follows:
∑ [
] (
)
(1)
where Q is the modularity index, m is the number of edges in the network, Aij is the number of
edges between node i and node j, ki and kj are the degree of nodes i and j respectively, and
δ(ci,cj) is a delta function that takes the value of 1 if i and j belong to the same community,
and zero otherwise. The term in brackets compares the number of edges between two
destinations with the number of edges we would expect if edges were distributed randomly in
the network, providing that the degree of each destination is not altered. Hence, the term in
brackets compares the actual number of relationships with a benchmark number of
relationship. If the number of edges between i and j is higher than the benchmark, these
destinations will form a community. The network will be partitioned in a number of
communities that maximizes the value of Q.
The procedure to determine the optimum number of communities is not trivial, as the
number of possible combinations of destinations rises exponentially with the number of
destinations, making the exhaustive comparison of all possibilities unfeasible. To overcome
this problem, different algorithms have been proposed to maximize modularity and identify
communities within a network. In this paper, we use the algorithm proposed by Blondel et al.
(2008). However, as pointed out by Fortunato and Berthelemy (2007), the modularity
maximization algorithm has a resolution-limit limitation, as it might aggregate small
communities within a broader community. In order to resolve this limitation they suggest
applying the maximization algorithm iteratively. First, the algorithm is applied on the whole
network. Second, the algorithm is applied only on each community identified in the first step.
The process stops once the algorithm does not find any further partition. In each step, it
should be checked that the number of edges within the communities identified by the
algorithm is larger than the expected number of edges. This iterative process has the
advantage of identifying hierarchies of communities. At the beginning, the algorithm
identifies few communities, characterized by a large number of members with only just above
the average connections between them. However, with each iteration, large communities are
fragmented into smaller communities characterized by stronger ties among members.
6
To avoid a too long iterative process and awkward relationships, we run the
community detection algorithm on a sample of destinations which are served, at least, by 50
exporters in 2002. In addition to that, exporters have to export, at least, to two different
destinations. After applying this filter, the number of nodes declines from 175 to 65
(representing 79% of total Mexican exports) in 2002.
Figure 2 displays the process of community-identification in Mexico. In the first
iteration, the algorithm identifies two big communities. The first community is composed by
countries located in the American continent, except USA and Canada; and the second
community by the rest of countries. When we apply the algorithm on the community of
countries located in the American continent, three final communities (shaded in yellow)
emerge. The first community is formed by South American countries: Argentina, Bolivia,
Chile, Colombia, Ecuador, Paraguay, Peru, Uruguay and Venezuela. The second community
is formed by Central American countries and three Caribbean countries: Belize, Costa Rica,
Cuba, Dominican Republic, El Salvador, Guatemala, Honduras, Nicaragua, Panama and
Puerto Rico. The third community is formed by Caribbean countries: Bahamas, Barbados,
Haiti, Jamaica and Trinidad and Tobago. When we apply again the modularity maximization
algorithm on the rest of countries community we get a fourth final community composed by
large developed countries: Canada, France, Germany, Great Britain, Italy, Spain and USA. If
we further iterate the group of remaining countries we end up with six additional final
communities. The fifth group is composed by small European countries: Austria, Belgium,
Denmark, Finland, Netherlands, Norway, Sweden and Switzerland. The sixth community
only includes three large Asian countries: India, Indonesia and Turkey. The seventh
community encompasses countries located in or near the Middle-East area: Egypt, Israel,
Saudi Arabia and United Arab Emirates, and two peripheral European Union countries:
Portugal and Greece. The eighth community is composed by two Eastern European countries:
Poland and Russia. The ninth community is formed mostly by Asian countries: China, HongKong, Japan, Korea, Malaysia, Philippines, Singapore, Taiwan and Thailand; countries
located in Oceania: Australia and New Zealand; and two emerging countries: Brazil and
South Africa. The final community is formed by two small European Union countries:
Hungary and Ireland.
We can observe that gravity variables, such as geographical location, play a role in the
formation of communities. For example, most of South American countries are located in the
same community and most of Central American countries are located in the same community.
However, we also observe that there are some countries that do not obey the geographical
7
rule. For example, Brazil does not belong to the community of South American countries, but
rather to a heterogeneous group that encompasses distant and emerging countries. We also
observe that Canada and the US do not form a North American community, but are integrated
in a broader large high-income countries' community.
These cases point out that there might be other reasons besides those captured by
gravity variables that might explain similarity between countries. As the network of Mexican
export-destinations is built upon the choices taken by Mexican exporters, each community
identified in the network is a revealed synthetic indicator of all the variables that might
determine the degree of similarity among export destinations. Hence, belonging to a
community might be a superior criterion to identify similarities between countries than
(extended) gravity measures.
What does a community capture besides the forces controlled by gravity variables? As
mentioned before, a community captures any variable that influences the degree of similarity
among countries that is difficult to observe or measure. For example, at the aggregate level,
the existence of large distribution chains that happen to be present in some countries, but not
in others, might explain why some exporters have a higher presence in a community of
countries. It might also be the case, as suggested before, that countries share similar
preferences for the exporter country products, and these similar preferences are not well
captured by extended gravity measures. In addition to that, following the trade models based
on firm heterogeneity (Melitz, 2003), a community can be considered as a set of countries that
require a similar threshold productivity for a foreign firm to obtain profits. The threshold is
high if fixed and variable costs of exporting are large or the size of the destination market is
small. For example, for Mexican exporters, large European Union countries belong to a
different community than small European Union countries because, due to differences in size,
the former demands a lower threshold productivity to obtain profits than the latter. As
variable costs of exporting are one of the determinants of the threshold productivity to obtain
profits, a destination might belong to one community for a country and to a different
community to another country. As long as gravity variables do not capture completely how
the combination of fixed costs, variable costs and demand factors determine threshold
productivities, the community variable might render a superior indicator, such that it still has
explanatory power besides these variables.
To test that community captures other variables besides those controlling for gravity
measures, we run a regression to determine to what extent gravitational variables explain the
variation in the probability of belonging to the same community. The gravitational variables
8
we use to explain the similarity between two destinations are distance, sharing a border,
speaking the same language, being located in the same region, belonging to the same income
per capita quintile, having a common colonizer, belonging to the same regional trade
agreement and the combined number of total migrants (number of citizens born in destination
i that live in destination j plus the number of citizens born in destination j that live in
destination i).5 As shown in Table 1, a larger distance reduces the probability of belonging to
the same community; while sharing a border, speaking the same language, being located in
the same region and having a similar income per capita level increase the probability of
belonging to the same community. The number of migrants born in one destination and living
in other destination also raises the probability of belonging to the same community. Finally,
sharing the same colonizer and belonging to the same regional trade agreement do not have a
significant impact on the probability of belonging to the same community. We can observe
that gravitational variables only explain 26% of the differences in the probability of belonging
to the same community in Mexico. These results confirm that, besides gravity measures,
community captures other variables that enhance the degree of similarity among countries.
2.3. Industry-specific communities.
As explained in the introductory section of the paper, one of the key advantages of the
network-based similarity measure versus extended-gravity measures is that the former can be
calculated at the industry level. In the extended gravity framework, for example, the marginal
distance from an incumbent destination to a new destination is the same irrespective of the
industry; however, in the network framework, the incumbent and the new destinations might
belong to the same community in some industries but not in others. To capture this
possibility, we classify exporters in one of seven industries: agriculture, chemicals, machinery
& transport equipment, metals, non-metallic minerals, paper and textiles. For each industry,
we identify the industry-specific destination-communities and the rest of industries
destination-communities. The limitation of this analysis is that the number of destinations at
the industry level is smaller than at the aggregate level. To keep a sufficient amount of
destinations we set a less stringent criterion to determine the sample used to identify
communities. In particular, we only exclude from the sample those destinations that are
served only by one exporter. The sample that meets this requirement and is common for the
seven industries is formed by 55 destinations. They represent 69% of Mexican exports in
5
See Appendix 1 for a description of the data sources.
9
2002. To identify the communities we follow the same procedure as the one used to identify
communities in the whole network.
Tables A2 in the Appendix display the identified communities in each of the seven
industries in Mexico. We should stress that results should be taken with care as we use highly
aggregated industries and few firms might be linking some marginal destinations, leading to
firm-specific communities. As shown in the figures, there are differences in the number of
communities among industries and countries. The highest number of communities is found in
chemicals, 10, and the lowest in paper, 8. We observe that there are some broad communities
present in most industries, such as a cluster of South American countries, a cluster for Central
American countries and a cluster for Caribbean countries. However, the size of these clusters
and its members vary from industry to industry. Moreover, in some cases, the communities of
South American and Central American countries are merged. Asian countries tend to stay in
the same community as other Asian countries; however, the members vary from industry to
industry, and sometimes the Asian countries form different communities. This is also the case
for large and high-income developed countries, and smaller and high-income European
countries.
To assess the similarity among communities across industries we calculate an adjusted
Rand index. This index, ranging between -1 and +1, calculates the fraction of destination pairs
that belong to the same community in two different industries.6 As shown in Table A3 of the
appendix, the adjusted rand indexes lie between 0.17 and 0.41. In most cases the correlation
across partitions is weak, confirming that destinations tend to belong to different communities
when examining different industries. In the last column of Table A3 in the Appendix, we
calculate the adjusted Rand index for the industry partitions and the rest of industries
partitions for each industry. We observe a weak correlation between partitions. These results
point out that the communities identified at the industry-level are different to those identified
for the whole set of exporters.
3. Can communities predict the expansion path of firms’ export destination portfolio?
So far we have calculated network-based communities from the entire web and
industry-specific webs of Mexican firms’ export destination portfolio. The rest of the paper
investigates whether our network-based communities helps to predict which countries will be
6
Specifically, the index calculates the fraction of correctly classified (respectively misclassified) elements to all
elements. The index is adjusted to ensure that the expected value of the index for two random partitions is zero
(Hubert and Arabie, 1985).
10
chosen by exporting firms that expand their destination portfolio. We begin by explaining the
empirical model used to study the determinants of the expansion of destination portfolio by
exporting firms. Next we describe sample of firms used to perform the empirical analysis.
3.1. Empirical model and econometric specification
Let yiGt =(yi1t, yi2t,…,yiKt) denote the vector of a exporting firm i’s current destination
portfolio G made of K countries. We want to examine the decision about where to export
when a firm expands the portfolio of destinations G. If destinations share common
characteristics, having served one destination might reduce the sunk cost of entering similar
destinations. Hence, previous export destinations might determine the export-path. Our
interest lies in the quantification of the effect of “similarity" between countries on the
probability of entering a new export destination. In particular we want to examine two types
of measures: those based on gravity-type indicators and those based on our network analysis.
We derive our econometric equation from a simple model of export participation into
specific foreign markets by profit-maximizing firms that produce one good in the local market
and sell part of the production abroad. There are several alternative markets and firms have to
decide which markets to export to. At any period, exporting firms have the choice of entering
into a number of markets if it did not export to those markets in the previous period. Let igt
be firm i’s profits from exporting to market g in year t. We assume that the expected profits of
exporting to country g by firm i is a linear function of factors affecting the destination choice,
(2)
where the vector
includes variables that measure the "mass" of information about
destination g that firm i might obtain from previous exporting experience in other
destinations,
is a vector of destination-specific constant terms and
is a random term
denoting the unobservable (by the researcher) unique profit advantage to the firm i from
selling in country g.
An exporting firm will choose to export to a particular country if she earns the highest
possible profit. Formally, the gth country is chosen by firm i as a new export destination
(
(omitting the subscript t) if
). If the firm-specific random terms are
independently distributed, each with a Type I extreme value distribution, McFadden (1974)
showed that the probability of a firm i to choose a destination g is
(
)
(
∑
)
(
)
(3)
11
where
is the population relative frequency of exporting to destination g. The estimates are
obtained by maximizing the likelihood function,
∏ ∏
. The model described above is
known as conditional logit model (CLM). It is easy empirically to generalize the CLM to the
case in which a firm can choose more than one destination every year.7
We create three sets of variables that capture the "mass" of information about
destination g that firms obtained from previous exporting experience. First, we use our
network-based community variable,
. This variable takes the value of one if the
new export market belongs to the same community of any of the firm’s previous destinations
and zero otherwise.8
As additional controls, we include gravity and extended gravity variables in order to
control for “observable” similarity features between destination markets. Similarity between
the domestic market (Mexico) and each new export markets is determined by standard gravity
measures, such as geodesic distance between the domestic and the new export market, sharing
a land border, common language, being members of the same regional trade agreement and
having a large number of migrants from each country living in the other country. We also use
GDP and GDP per capita as a proxy for the attractiveness of the new export market.
For extended gravity variables, we control for distance, border, language, regional
trade agreement, migration, income level and geographical region. Then the variable
characterizes the countries' geographical relationship to prior export destinations of the firm.
It takes the value of one if the new export destination capital city is less than a certain number
of kilometers away from the capital city of any of the previous destinations and zero
otherwise. In the benchmark analysis, we use a 1,500 km. radius. We also proxy for the
geographical links between countries using a common border dummy variable,
,
which takes the value of one if the new export destination shares a land-border with any
previous destination and zero otherwise; and
, which takes the value of one if the new
export destination is in the same region of any of the previous destinations and zero
7
Notice that CLM does not allow the inclusion of explanatory variables that are not directly related to the
choices. In our case, it means that we cannot estimate a single parameter to capture the impact of firm-specific
characteristics on the firm’s probability of exporting to a particular destination. Another potential limitation of
CLM is the risk of violation of the Independence Irrelevant Alternatives (IIA) assumption. In the sensitivity
analyses section we estimate equation (3) with alternative models such as nested logit and mixed logit, that
relaxes the IIA assumption and allow incorporating firm-level characteristics in the model.
8
Notice that we constructed our “stock” community variable with 2002 data, before new exporters start
expanding their “flow” of destinations since 2004 onwards.
12
otherwise.9 The regional dummy variable also controls for the existence of natural trade blocs
in international trade (Frankel et al., 1995). We also consider cultural closeness measures such
as common language between export destinations. Specifically, the variable
takes
the value of one if the new export destination speaks the same language of any of the previous
destinations and zero otherwise. We also proxy for economic proximity, controlling the
presence of any previous export destination located in the same income quintile of the new
export market. We also consider other variables that might enhance the proximity between
destinations, such as migration flows and belonging to the same regional trade agreement.
Specifically, the variable
takes the value of one if any previous destination has at least
100 immigrants and 100 emigrants in the new export destination and zero otherwise. The
variable
takes the value of one if the new export destination belongs to the same
regional trade agreement of any of the previous destinations and zero otherwise.
3.2. Data
Our sample consists of all (6026) Mexican exporting firms that internationalized
between 2003 and 2007 and carried on exporting until 2009. We called these firms “new
regular exporters”.10 There are some interesting features that explain why we have chosen
them. First, we know their entire export portfolio since we know they started exporting.
Second, they account for a significant share of Mexican exports (21% of total firms, 23% of
all transactions and 21% of all value of exports in 2009). Third, as Table 2 shows, they exhibit
dynamics of export destination portfolio similar to the old regular exporters. Considering
year-to-year changes, the largest percentage of firms (about 60 percent) does not modify the
export destination portfolio. Every year some firms only enter into new destinations (15%),
others opt for only exiting (9%) and, finally, others decide to enter and exit simultaneously
(17%). The last column in Table 2 shows the transitions in the export destination portfolio of
the old regular exporters (firms exporting before 2003) in the period 2007-2008. They are
almost the same as those of the new regular exporters.
Table 3 presents a summary of our dependent variable: the number of new destinations
per year served by a typical new regular exporter. The percentage of firm-year pairs that take
a value of one is 56%, that is, the majority of firms in our sample that decide to expand their
9
We use the seven major regions identified by the World Bank: East Asia & Polynesia, Europe & Central Asia,
Latin-American & Caribbean, Middle East & North Africa, South Asia, and Sub-Saharan Africa.
10
Strictly speaking, we know that a new regular exporter did not export in 2000, 2001 and 2002. In our set of
new regular exporters, 757 started to export in 2003, 948 in 2004, 1110 in 2005, 1283 in 2006 and 1928 in 2007.
13
destination portfolio enter a single new destination. The number of firms that expand their
destination portfolio in more than six destinations is very small (less than 3% of the firmyear).
Next we show the distribution of destinations across communities. The number of
communities served by a typical new regular exporter each year is one: 75% of new regular
exporters only serve one community, 14% of new regular exporters serve two communities,
6% of new regular exporters serve three communities and 5% of new regular exporters serve
four or more communities. As the US is the most important destination for Mexican firms, a
very large percentage of new regular exporters (84%) serve the community in which the US is
integrated 11 ; 22% serve the community of Central American countries, 13% serve the
community of South American countries and another 13% the community of Asian countries.
The community of small European countries is served by 5% of new regular exporters; the
community of Middle East countries, the community formed by India, Indonesia and Turkey,
and the community of Caribbean countries are only served by 2% of the new regular
exporters. The rest of communities are served by very few new regular exporters.
4. Estimation results
In this section we present the results from the econometric analyses. First, we
investigate whether communities identified in the whole network of Mexican exporters
determine the export path of new exporters. We also analyze whether industry-specific
destination-communities have a larger role in determining the export path than general
destination-communities. Second, we perform a set of sensitivity analyses to test the
robustness of our results.
4.1. Main results
Table 4 reports estimates of the conditional logit model allowing for simultaneous
exports to multiple destinations. In specification (1) we estimate the model with community
as the only independent variable. The communities used in the benchmark analysis are the
final communities identified in Figure 2. The community coefficient has a large positive value
and is strongly statistically significant. The transformation of the community coefficients into
odds-ratios provides an easy way to interpret economically the estimates. For Mexico, the
probability of choosing a new destination rises by 474% (=exp(1.557)) if it belongs to the
same community of any of the firm’s previous destinations.
11
79% of new regular exporters served the US market in the period 2003-2009.
14
Specification (2) introduces gravitational variables, such as GDP, GDP per capita,
distance, common border, common language, belonging to the same regional trade agreement
and migration stocks to proxy the similarity between the domestic (Mexico) and the new
export market, and the attractiveness of the new export market. There is a reduction in the size
of the positive community coefficient and still is strongly statistically significant. According
to the new coefficients, once we control for gravity measures, the probability of choosing a
new destination rises by 303% in Mexico if it belongs to the same community of any of the
firm’s previous destinations. Regarding gravitational variables, the larger the size of the new
export market the higher the probability of choosing that market as a new export destination;
and the larger the bilateral distance the lower the probability of selecting the new export
destination. Speaking the same language, having a land border, belonging to the same trade
agreement and a higher number of migrants raise the probability of choosing the new export
destination. In contrast, the larger the new export market's income per capita the lower the
probability of selecting that market.
Specification (3) also controls for extended-gravity measures and, again, the positive
value of the community coefficient is further reduced: once we control for gravity and
extended gravity measures, the probability of choosing a new export destination rises by
193% if it belongs to the same community of any of the firm’s previous destinations.
Regarding extended gravity variables, having a previous export destination within
1,500 kilometers radius of the new destination raises the probability of selecting this
destination. We also find that a new export destination has a larger probability of being
chosen if there are previous destinations that speak the same language as this new destination,
share a border, are located in the same income-quintile and region, and have sizable bilateral
migration flows. In contrast, probability of being chosen is not affected when the new
destination belongs to the same regional trade agreement of any previous destination.
To confirm that the community variable enhances the model’s predictive capacity, we
calculate the percentage of cases in which the observation with the highest predicted
probability by the conditional logit model corresponds to the new destination selected by the
exporter. We compare this percentage in a model that includes gravity and extended gravity
variables and a model that also includes the community variable. We find that adding the
community variable improves the accuracy of predictions by almost 20%.
Finally, specification (4) introduces destination-specific fixed effects. These fixed
effects preclude the estimation of time-invariant gravity variables so, for the sake of clarity,
we remove all gravity variables from the estimation. The community coefficients remain
15
positive and statistically significant at the 1 percent level. Once we control for gravity,
extended-gravity and destination-specific effects, the probability of choosing a new export
destination rises by 180%, almost a 3-fold increase, if it belongs to the same community of
any of its previous destinations.
Table 5 presents the results of estimating the model with industry-specific
communities and rest of industries communities. Both coefficients are positive and strongly
statistically significant. As we expected, the industry-specific coefficient is larger than the rest
of industries community coefficient. However, the differences are small and we cannot reject
the null hypothesis that both coefficients are equal. Hence, we conclude that industry-specific
communities and rest of industries communities play similar roles in determining exporters'
export-path.
4.2. Sensitivity analyses
We perform a series of sensitivity analyses to assess the robustness of our results.
First, we want to confirm that the communities identified by the network-analysis algorithm
do not have a strong explanatory capacity by chance. To rule out this possibility we assign
destinations to communities randomly. To carry out this exercise we assume that the number
of communities and the number of members within each community is the same as in the
benchmark estimations. We perform the exercise 50 times; each time, once the random
communities are generated we run the same model as the one in Table 4 - specification (4). In
all 50 estimations the community coefficient never was positive and statistically significant.
These results point out that the communities generated by the network-analysis algorithm do
not exert an influence on export dynamics by chance. On the contrary, it confirms that
belonging to a community is a very important determinant of the evolution of the export path.
Second, we analyze whether the community coefficient is robust to the use of a larger
sample to identify communities. We expand the sample including the years 2000 and 2001.
To avoid marginal destinations, we exclude from the sample the destinations with less than 50
exporters during the period 2000-2002. The longer period and a less stringent threshold to
admit a destination raises the number of destinations to 90 (65 previously). After applying
iteratively the modularity maximization algorithm, we identify 12 communities. Compared to
the sample used in the benchmark analyses, the number of communities rises in two (see
Tables A4 in the Appendix). Table 6 presents the results of estimating the model with the new
samples. The community coefficient remains positive and strongly statistically significant.
Compared with the benchmark estimation, the value of the community coefficient increases.
16
According to the new estimates the probability of selecting a new export destination rises by
200% if it belongs to the same community of any of the firm’s previous destinations.
Third, we analyze whether the community coefficients are robust to different extended
distance variables. The variable used in the benchmark analyses is whether there is a previous
export destination within a 1,500 kilometers radius of the new export destination. We reestimate the model with a shorter distance: 500 kilometers, and a larger distance 3,000
kilometers. As shown in columns 2-3 in Table 6, the community coefficient is robust to the
alternative extended-gravity distance radius.
Fourth, we use a more stringent threshold to determine whether a firm is a new
exporter. Now, we define a firm as a new exporter if it does not export in 2000, 2001, 2002
and 2003. As shown in column 4 of Table 6, the community coefficient remains positive and
statistically significant and is similar to those reported in the benchmark estimation.
Fifth, we convert the community and the extended gravity variables from discrete
variables to count variables. For example, now, the community coefficient is the number of
previous destinations that belong to the same community as the new destination. As shown in
column 5 of Table 6, now the community coefficient reduces its positive value, but remains
strongly statistically significant: a one unit increase in the number of destinations previous
served by the exporter within the community rises the probability of choosing a new
destination within the community by 128%.
Sixth, we analyze whether results for Mexico are robust to excluding all transactions
with the US. As explained previously, 79% of new regular exporters in Mexico have the US
in their portfolio and, hence, most of new regular exporters serve the community in which the
US is integrated: 84%. To ensure that results are not driven by the large percentage of
exporters having the US as destination, we remove all export transactions to the US from the
new regular exporters' database. As shown in column 6 of Table 6, the community coefficient
is very similar to the one reported in the benchmark analysis table so our estimations are not
driven by the large percentage of new regular exporters that serve the US market.
Seventh, the results presented in the benchmark estimation use the final communities
identified in Figure 2. In order to test the robustness of our results we also estimate the model
with communities identified with fewer iterations of the modularity maximization algorithm.
A shown in Figure 2, the maximum number of iterations needed to arrive to a final destination
is five. So we can use the communities identified after four iterations, three iterations, two
iterations and one iteration. As we reduce the number of iterations, the number of
communities is also reduced. With a lower number of communities, clusters have a larger
17
number of members but we expect similarities between members to be lighter. These effects
might drive the community coefficient in opposite directions. On the one hand, as there are
more destinations within a community, exporters will have a larger probability of choosing a
destination within a community. On the other hand, as similarities within the community are
lighter, exporters will have a lower probability of choosing a destination belonging to the
community.
Table 7 presents the results of the estimations for different community hierarchies. For
comparison we also reproduce the results when estimating the model with the final
communities (column 1). The community coefficient is always positive and statistically
significant; the community coefficient only drops substantially when we only use one
iteration.
Eighth, the (fixed-effects) conditional logit model imposes a strong restriction: IIA.
The IIA states that the ratio of the probabilities of choosing a new export market only depends
on the attributes of the two destinations, and is independent on the characteristics of other
destinations. However, this restriction fails if some destinations have some (unobserved)
common characteristics, making substitution among them easier. If the IIA restriction does
not hold the conditional logit model leads to biased estimates. To address this limitation, we
estimate alternative logit models that relax the IIA assumption: the nested-logit model and the
mixed logit model. In the nested model, alternatives can be separated, at least, in two main
groups. Within each group the IIA assumption holds, but across groups the IIA assumption
does not need to hold. To implement the nested logit model we should determine a nesting
criterion. We start assuming that firms decide, first, what major world region they want to sell
to and, second, they select the destination within that region. We also used alternative nesting
criteria, such as dividing destinations into close markets and distant markets, which no major
changes in results. To estimate the nested logit model we assign destinations to one of the
seven major regions defined by the World Bank.12 The results after estimating the nested logit
model are reported in Table 8. The community coefficient remains positive and highly
statistically significant. As shown at the bottom of the table the LR test rejects the IIA
assumption, validating the nesting of destinations by major regions.
A limitation of the nested logit model is that it introduces a very rigid substitution
structure. For example, in our previous exercise a destination can only belong to one region.
The mixed logit model overcomes this rigidity by allowing for variation in the coefficients
12
See footnote 10.
18
across firms, unrestricted substitution patterns, and correlation in unobserved factors over
time (Train, 2003). The last two columns of Table 8 present the results of estimating the
mixed-logit model. One column presents the average value of each coefficient and the other
column presents the average standard deviation of each coefficient. 13 The community
coefficient remains positive and highly statistically significant. Looking to the standard
deviation coefficient, we observe, as well, that there is a large heterogeneity in the impact of
community across firms.
To sum up, the sensitivity analyses show that the positive and significant contribution
of belonging to a community in determining the dynamics of firms’ new export destinations is
robust to the use of different samples and econometric specifications. We also show that this
positive relation does not arise randomly.
5. Conclusions
How do exporters choose new export destinations? While there are many factors that
are important for this decision, an empirical regularity strikes out: firms tend to choose new
export markets that are similar to their prior export destinations. Network analysis, through
the community-detection algorithm, provides a tool to identify destinations that share a
common set of characteristics. This measure has three advantages over the gravity measures.
First, it is a revealed measure and, hence, encapsulates all the observable and non-observable
factors that may influence the degree of similarity among destinations. Second, it is a countryspecific measure. Third, it allows the calculation of industry-specific similarity measures.
We apply this methodology to the web of Mexican exporters' destinations and find that
there are ten communities in the Mexican web (of 65 countries). Next we show that belonging
to the same community of a previous export destination exerts a strong influence on the
dynamics of the exporter. In particular, the probability of choosing a new destination
multiplies by three if it belongs to the same community of any of firm’s previous destinations.
We show that the strong predictive capacity of the network-based similarity measure remains
once we control for gravity measures, and improves the accuracy of the model up to 20%. We
also show that industry-specific community of destinations and general community of
destinations exert a similar influence on the dynamics of Mexican exporters' destinationportfolio. Results are robust to different specifications and samples.
13
In contrast to the conditional logit model, the nested and mixed logit models do not allow firms to choose
more than one new destination per year. Hence, when estimating these latter models, we narrow the sample to
exporters than only choose one new destination per year.
19
Appendix. Data sources.
Our firm-level data comes from transaction-level customs data on the universe of
Mexican exporters over the period 2000-2009. The source for the data is detailed in the
Annex of Cebeci, Fernandes, Freund and Pierola (2012) and the data was collected by the
Trade and Integration Unit of the World Bank Research Department, as part of their efforts to
build the Exporter Dynamics Database. In the Mexican dataset the firm identification code
changes from 2007 onwards. For the year 2007 we have data with the old firm classification
and with the new firm classification. Matching firm-level country and HS 6-digit specific
records we can establish a correspondence between the old firm classification and the new
firm classification for firms that exported in 2007. For the rest of firms that exported in 2008
and/or 2009, we cannot know whether they are new exporters or they are firms that exported
in the 2000-2006 period. Since we use data on year 2002 for the network analysis and the
sample of entrants that export at three consecutive years for the analysis of the dynamics of
destination portfolio, this problem in the raw data does not affect our analysis. Table A1 in
the Appendix display information on the number of trading firms, number of transactions and
value (in millions US$) in the 2000-2009 period.
Data for the construction of the gravity and extended-gravity measures come from
different sources. Data on income and population are taken from World Bank (2012).
Distance, contiguity, common language, colonial relationship and same continent are obtained
from Head et al. (2010). Bilateral migration stocks in 2000 are from Özden et al. (2011) and
common membership in a regional trade agreement (RTA) in 2002 is obtained from de Sousa
et al. (2012). The web links for databases open to the public are: World Development
Indicators:
gravity
http://data.worldbank.org/data-catalog/world-development-indicators;
dataset:
http://www.cepii.fr/anglaisgraph/bdd/gravity.asp;
RTA
CEPII
database:
http://kellogg.nd.edu/faculty/fellows/bergstrand.shtml; World Bilateral Migration database:
http://data.worldbank.org/data-catalog/global-bilateral-migration-database.
References
Albornoz, F., Calvo-Pardo, H., Corcos, G., and Ornelas, E. (2012). "Sequential exporting",
Journal of International Economics, 88, 1, 17-31
Besedes, T. and Prusa, Th. (2011), “The Role of Extensive and Intensive Margins and Export
Growth”, Journal of Development Economics, 96(2), 371-379
Blondel, V. D., Guillaume, J.L., Lambiotte, R. and Lefebvre, E. (2008). "Fast unfolding of
communities in large networks", Journal of Statistical Mechanics: Theory and
Experiment, October, P10008.
20
Cebeci, T., Fernandes, A.M., Freund, C. and Pierola, M.D. (2012). “Exporter Dynamics
Database”, World Bank Policy Research Paper 6229.
Cadot, O., Iacovone, L., Pierola, M.D., Rauch, F. (2013), “Success and failure of African
exporters”, Journal of Development Economics, 101, 284-296.
Chaney, T. (2011). "The network structure of international trade", NBER Working Paper
16753.
De Benedictis, L. and Tajoli, L. (2011). "The World Trade Network", The World Economy,
34, 8, 1417-1454.
de Sousa, J., Mayer, T. and Zignano, S. (2012). “Market access in global and regional trade”,
Regional Science and Urban Economics, 42, 6, 1037-1052.
Defever, F., Heid, B. and Larch, M. (2011). "Spatial exporters", CEP Discussion P. No. 1100.
Evenett, S.J. and Venables, A. J. (2002). "Export Growth in Developing Countries: Market
Entry and Bilateral Trade Flows", University of Bern Working Paper, mimeo.
Fortunato, S. and Berthélemy, M. (2007). "Resolution limit in community detection",
Proceedings of the National Academy of Sciences, 104, 1, 36-41.
Frankel, J., Stein, E. and Wei, S. (1995). “Trading Blocs and the Americas: The Natural, the
Unnatural, and the Super-Natural”, Journal of Development Economics, 47, 1, 61-95.
Hidalgo, C.A., B. Klinger, A.L. Barabási and R. Hausmann, R. (2007). "The Product Space
Conditions the Development of Nations", Science, 317, 5837, 482-487.
Hubert, L. and Arabie, P. (1985). "Comparing partitions", Journal of Classification, 2, 1,
193–218.
Kali, R. and Reyes, J. (2007). "The architecture of globalization: a network approach to
international economic integration", Journal of International Business Studies, 38, 4,
595-620.
Kali, R., Reyes, J., McGee, J. and Shirrell, S. (2013). "Growth networks", Journal of
Development Economics, 101, 216-227.
Lawless, M. (2011). "Marginal Distance: Does Export Experience Reduce Firm Trade
Costs?", Central Bank of Ireland Technical Working Paper 02/RT/11.
McFadden, D. (1974). "Conditional Logit Analysis of Qualitative Choice Behavior", in P.
Zarembka (eds.), Frontiers in Econometrics, New York: Academic Press, 105–142.
Morales, E., Sheu, G., and Zahler, A. (2011). "Gravity and extended gravity: Estimating a
structural model of export entry", unpublished paper.
Newman, M.E. (2006). "Modularity and community structure in networks", Proceedings of
the National Academy of Sciences, 103, 23, 8577-8582
Nguyen, D. X. (2012). "Demand uncertainty: Exporting delays and exporting failures",
Journal of International Economics, 86, 2, 336-344.
Özden, C., Parsons, C.R., Schiff, M. and Walmsley, T.L. (2011). “Where on Earth is
Everybody? The Evolution of Global Bilateral Migration 1960-2000”, World Bank
Policy Research Working Paper 5709.
Train, K.E. (2003). Discrete Choice Methods with Simulation, Cambridge University Press,
Cambridge.
21
Figure 1. The network of Mexican firms export destinations, 2002
22
Figure 2. Community detection process in the network of Mexican exporters' destinations
ARG, BHS, BLZ, BOL, BRB,
CHL, COL, CRI, CUB,
DOM, ECU, GTM, HND,
HTI, JAM, NIC, PAN, PER,
PRI, PRY, SLV, TTO, VEN
ARG, BOL, CHL, COL, ECU, PER, PRY,
URY, VEN
BLZ, CRI, CUB, DOM, GTM, HND, NIC,
PAN, PRI, SLV
AUT, BEL, CHE,
DNK, FIN, NLD,
NOR, SWE
BHS, BRB, HTI, JAM,
TTO
65 top
destinations
CAN, DEU, ESP,
FRA, GBR, ITA,
USA
ARE, AUS, AUT, BEL, BRA,
CAN, CHE, CHN, DEU,
DNK, EGY, ESP, FIN, FRA,
GBR, GRC, HKG, HUN,
IDN, IND, IRL, ISR, ITA,
JPN, KOR, MYS, NLD, NOR,
NZL, PHL, POL, PRT, RUS,
SAU, SGP, SWE, THA, TUR,
TWN, USA, ZAF
ARE, AUS, AUT, BEL,
BRA, CHE, CHN, DNK,
EGY, FIN, GRC, HKG,
HUN, IDN, IND, IRL, ISR,
JPN, KOR, MYS, NLD,
NOR, NZL, PHL, POL,
PRT, RUS, SAU, SGP,
SWE, THA, TUR, TWN,
ZAF
ARE, AUT,
BEL, CHE,
DNK, EGY,
FIN, GRC, ISR,
NLD, NOR,
POL, PRT, RUS,
SAU, SWE
AUS, BRA,
CHN, HKG,
HUN, IDN, IND,
IRL, JPN, KOR,
MYS, NZL,
PHL, SGP, THA,
TUR, TWN,
ZAF
ARE, EGY,
GRC, ISR,
PRT, SAU
ARE, EGY, GRC,
ISR, POL, PRT,
RUS, SAU
IDN, IND,
TUR
AUS, BRA, CHN,
HKG, HUN, IRL, JPN,
KOR, MYS, NZL,
PHL, SGP, THA,
TWN, ZAF
POL, RUS
AUS, BRA,
CHN, HKG,
JPN, KOR,
MYS, NZL,
PHL, SGP,
THA, TWN,
ZAF
HUN, IRL
23
Table 1. Regression results of the probability of belonging to the same community on
gravity measures
Distance (Ln)
-0.044*** (0.009)
Border
0.133*** (0.028)
Language
0.135*** (0.013)
Region
0.177*** (0.016)
Income
0.034*** (0.009)
Common colonizer
Regional trade agreement
0.034 (0.027)
-0.020 (0.016)
Migration flows (Ln)
0.011*** (0.001)
Constant
0.359*** (0.085)
R-square
0.26
Observations
4,160
Note: Linear probabilistic model. Standard deviations in parentheses. ***, ** statistically significant at
1% and 5% respectively.
24
Table 2. Changes in export destination portfolio of Mexican “new regular exporters”.
(1)
Type of firm
Years exporting
# new regular exporters
# firms in each period
Changes in country portfolio (%)
Only entries
Only exits
Simultaneous entry & exit
same entries and exits
entries>exits
entries < exits
No change in portfolio
(2)
(3)
(4)
(5)
New exporting firms since 2003 that do not stop exporting
03-04
04-05
05-06
06-07
07-08
757
948
1110
1283
1928
757
1705
2815
4098
6026
(6)
Regular
exporters
2000-2009
07-08
5697
21
4
17
6
15
8
14
8
15
9
16
7
5
8
1
61
6
7
3
62
5
7
4
61
6
7
4
61
6
7
4
60
6
7
3
61
Source: Own elaboration using Census of Exporting Mexican Firms, 2000-2009.
25
Table 3. Dependent variable. Number of new export destinations per firm-year.
# entries per
firm-year pair
1
2
3
4
5
6
7
8
9
10
11 or more
Total
Frequency
4104
1524
701
363
212
129
54
49
41
38
32
7247
Cum. Freq.
56.63
21.03
9.67
5.01
2.93
1.78
0.75
0.68
0.57
0.52
0.44
100.00
26
Table 4. Main results. Conditional logit estimations. Baseline results. (Dependent
variable: Choice of new export destination)
Specification
Community
GDP
GDP pc
Distance
Border
Language
RTA
Migrants
I_distance 1500
I_border
I_language
I_RTA
I_income
I_migration
I_region
Country
dummies
Observations
Nº of firms
Nº of countries
R2
(1)
1.557*** (0.025)
(2)
1.108*** (0.027)
0.487*** (0.009)
-0.187*** (0.012)
-0.995*** (0.026)
0.265*** (0.040)
0.444*** (0.031)
0.107*** (0.024)
0.034*** (0.004)
No
No
(3)
0.662*** (0.013)
0.530*** (0.009)
-0.212*** (0.012)
-0.827*** (0.026)
0.573*** (0.042)
0.484*** (0.032)
0.154*** (0.024)
0.017*** (0.005)
0.069** (0.024)
0.426*** (0.030)
0.338*** (0.026)
-0.039 (0.031)
0.130*** (0.031)
0.810*** (0.058)
0.425*** (0.037)
No
446,161
3,339
65
0.07
446,161
3,339
65
0.16
446,161
3,339
65
0.18
(4)
0.587*** (0.031)
446,161
3,339
65
0.21
0.334*** (0.030)
0.436*** (0.030)
0.331*** (0.037)
0.216*** (0.034)
0.174*** (0.030)
0.503*** (0.059)
0.523*** (0.039)
Yes
Note: Clustered standard errors at the firm-level in parentheses. ***, ** significant at 1 percent and 5
percent respectively.
27
Table 5. Estimations with sector-specific communities
Dependent variable: Choice of new export destination.
Mexico
I _Community sector
I _Community no-sector
I _distance 1500
I _border
I _language
I _RTA
I _income
I _migration
I _region
Observations
Nº of firms
Nº of countries
R2
0.447*** (0.030)
0.370 *** (0.036)
0.313*** (0.030)
0.379*** (0.031)
0.363*** (0.038)
0.208*** (0.035)
0.139*** (0.031)
0.465*** (0.062)
0.583*** (0.040)
407,742
3,440
55
0.21
Note: All estimations include destination-specific fixed effects. Clustered standard errors at the firm-level in
parentheses. *** significant at 1 percent.
28
Table 6. Sensitivity Analyses I (Dependent variable: Choice of new export destination)
Specification
Community
I_distance 1500
(1)
Larger
sample
0.693***
(0.032)
0.304***
(0.030)
D_distance 500
(2)
N_distance
500
0.587***
(0.031)
D_language
D_RTA
D_income
D_migration
D_region
0.595***
(0.031)
(4)
New
exporters
(5)
Count
(6)
No US
0.560***
(0.034)
0.325***
(0.033)
0.250***
(0.012)
0.070***
(0.020)
0.572***
(0.031)
0.299***
(0.030)
0.390***
(0.034)
0.379***
(0.041)
0.250***
(0.038)
0.175***
(0.034)
0.483***
(0.064)
0.541***
(0.044)
0.294***
(0.022)
0.036***
(0.009)
0.027***
(0.005)
0.066***
(0.010)
0.007
(0.009)
0.071***
(0.011)
0.444***
(0.031)
0.397***
(0.040)
0.248***
(0.038)
0.180***
(0.032)
0.477***
(0.061)
0.517***
(0.042)
0.266***
(0.046)
D_distance 3000
D_border
(3)
N_distance
3000
0.400***
(0.030)
0.037***
(0.009)
0.189***
(0.034)
0.150***
(0.030)
0.536***
(0.059)
0.581***
(0.039)
0.347***
(0.033)
0.336***
(0.037)
0.233***
(0.034)
0.155***
(0.030)
0.470***
(0.058)
0.619***
(0.038)
0.090***
(0.034)
0.434***
(0.030)
0.340***
(0.037)
0.234***
(0.034)
0.173***
(0.030)
0.490***
(0.058)
0.601***
(0.039)
Observations
468,883
446,161
446,161
357,715
446,161
333,126
Nº of firms
3,328
3,320
3,320
2,826
3,320
2,750
Nº of countries
90
65
65
65
65
64
R2
0.23
0.20
0.20
0.21
0.19
0.18
Note: All regressions include destination-specific fixed-effects. Clustered standard errors at the firm-level in
parentheses. ***, ** significant at 1 percent and 5 percent respectively.
29
Table 7. Sensitivity Analyses II: Estimations for different community hierarchies
(Dependent variable: Choice of new export destination).
Community type
Final
Community
0.587***
(0.031)
0.334***
(0.030)
0.436***
(0.030)
0.331***
(0.037)
0.216***
(0.034)
0.174***
(0.030)
0.503***
(0.059)
0.523***
(0.039)
446,161
3,339
65
0.21
D_distance 1500
D_border
D_language
D_RTA
D_income
D_migration
D_region
Observations
Nº of firms
Nº of countries
R2
4 iterations
0.592***
(0.030)
0.332***
(0.030)
0.435***
(0.030)
0.332***
(0.037)
0.215***
(0.034)
0.171***
(0.030)
0.498***
(0.058)
0.525***
(0.039)
446,161
3,339
65
0.21
3 iterations
0.612***
(0.031)
0.335***
(0.030)
0.436***
(0.030)
0.339***
(0.037)
0.204***
(0.034)
0.166***
(0.031)
0.477***
(0.059)
0.524***
(0.039)
446,161
3,339
65
0.21
2 iterations
0.625***
(0.030)
0.335***
(0.030)
0.434***
(0.030)
0.332***
(0.037)
0.199***
(0.034)
0.157***
(0.031)
0.473***
(0.058)
0.556***
(0.039)
446,161
3,339
65
0.21
1 iteration
.
0.473***
(0.044)
0.366***
(0.030)
0.493***
(0.030)
0.437***
(0.036)
0.320***
(0.033)
0.294***
(0.030)
0.395***
(0.058)
0.608***
(0.039)
446,161
3,339
65
0.20
Note: All estimations include destination-specific fixed effects. Clustered standard errors at the firm-level in
parentheses. ***, **, * significant at 1 percent and 5 percent respectively.
Number of observations=446161. Number of firms: 3339. Number of countries: 65.
30
Table 8. Sensitivity Analyses III: Nested-Logit and Mixed -Logit model estimations
(Dependent variable: Choice of new export destination).
Specification
Community
D_distance 1500
D_border
D_language
D_RTA
D_income
D_migration
D_region
LR-Test for IIA (phi value)
Nested-Logit
0.760*** (0.061)
0.561*** (0.086)
0.932*** (0.093)
0.587*** (0.073)
0.226*** (0.068)
0.262*** (0.063)
0.801*** (0.131)
Mixed-Logit
Mixed-logit
Mean
Standard deviation
0.613*** (0.054)
0.959*** (0.113)
0.193** (0.083)
0.544*** (0.197)
0.493*** (0.067)
0.599*** (0.164)
0.340*** (0.048)
0.368* (0.191)
-0.094* (0.048)
0.094 (0.136)
0.210*** (0.056)
0.708*** (0.133)
1.051*** (0.181)
0.742** (0.303)
0.392*** (0.066)
0.561*** (0.153)
0.0000
Note: The mixed-logit estimation also includes gravity-type controls. Standard deviation in parentheses.
***, **, * significant at 1 percent, 5 percent and 10 percent respectively. Number of
observations=250157. Number of firms: 2625. Number of countries: 65.
31
Table A1. Mexican exporters database (2000-2009)
Number of exporting firms
Number of transactions
Value of exports (current US$ million)
Regular exporters (N=5697) (% firms)
% total transactions
% value of total exports
2000
35.509
183.586
165.974
2001
34.318
181.343
158.539
2002
31.592
171.029
160.669
2003
30.420
161.899
164.941
2004
30.441
167.905
187.736
2005
30.984
173.620
213.902
2006
30.171
174.308
249.510
2007
30.283
187.267
270.776
2008
29.796
190.584
290.160
2009
28.690
183.036
228.728
16%
35%
60%
17%
38%
64%
18%
40%
65%
19%
40%
66%
19%
43%
66%
18%
46%
66%
19%
46%
67%
19%
46%
67%
19%
46%
68%
20%
46%
69%
757
2%
3.207
2%
30.890
19%
948
3%
3.907
2%
29.351
16%
1.110
4%
5.081
3%
28.498
13%
1.283
4%
5.145
3%
33.411
13%
1.928
6%
7.177
4%
62.596
23%
165.396
187.980
214.207
249.961
271.821
Regular new exporters
Number
(% in total firms)
Transactions
(% in total transactions)
Value of exports
(% in total exports)
Memorandum
Merchandise exports (Source: WDI)
166.367
158.547
160.682
6.026
21%
41.571
23%
47.805
21%
291.265
229.712
Note: The term "firm" refers to any individual operator that makes a transaction in a year. The dataset contains all the transactions with a value above 1,500US$. The area in
dark grey refers to the number of firms, number of transactions and value of exports in 2009 of the new regular exporters, that is, firms that started exporting after 2003 and
did not stop exporting until 2009 (our Mexican firms sample).
32
Table A2. Communities by industries. Mexico 2002
Agriculture
Chemicals
ARG
BOL
BRA
CHL
COL
DOM
ECU
PER
PRI
VEN
BHS
BLZ
BRB
CUB
HTI
JAM
TTO
URY
CRI
GTM
HND
NIC
PAN
SLV
CHN
KOR
PHL
THA
HKG
IDN
NZL
SGP
TWN
AUS
BEL
CAN
CHE
DEU
ESP
GRA
GBR
ITA
JPN
NLD
SWE
USA
ARE
GRC
MYS
PRT
SAU
AUT
FIN
HUN
ZAF
BOL
CHL
COL
CUB
ECU
PAN
PER
VEN
BHS
BRB
JAM
TTO
CRI
DOM
GTM
HND
NIC
SLV
ARE
CHN
HKG
IDN
KOR
THA
TWN
AUS
GRC
MYS
NLZ
PHL
SGP
DEU
ESP
FRA
GBR
ITA
JPN
BLZ
CAN
HTI
PRI
USA
FIN
HUN
SAU
ARG
BRA
PRT
URY
ZAF
AUT
BEL
CHE
NLD
SWE
Machinery &
Transport equipment
CHL
COL
CRI
CUB
DOM
ECU
GTM
HND
NIC
PAN
PER
SLV
USA
VEN
BHS
BLZ
BOL
BRB
HTI
JAM
PRI
TTO
ARG
IDN
MYS
NZL
PHL
SGP
THA
TWN
URY
ZAF
AUS
BEL
BRA
CAN
CHN
DEU
ESP
FRA
GBR
HKG
ITA
JPN
KOR
NLD
ARE
FIN
GRC
SAU
AUT
CHE
HUN
PRT
SWE
Metals
BOL
CHL
COL
ECU
PER
VEN
BHS
BLZ
BRB
HTI
JAM
TTO
CRI
CUB
DOM
GTM
HND
NIC
PAN
PRI
SLV
HKG
HUN
MYS
NLD
SGP
THA
TWN
AUS
AUT
BEL
BRA
CHN
GBR
JPN
KOR
ZAF
CAN
CHE
DEU
ESP
FRA
ITA
USA
ARE
SAU
ARG
IDN
PHL
URY
FIN
GRC
NZL
PRT
SWE
Non-metallic
minerals
ARG
BRA
CHL
COL
ECU
ESP
PER
VEN
BLZ
GRC
HTI
JAM
TTO
BHS
CHE
CRI
CUB
DOM
GTM
HND
NIC
PAN
PRI
SLV
USA
ARE
IDN
KOR
THA
TWN
CHN
HKG
JPN
PHL
ZAF
CAN
FRA
GBR
ITA
SAU
SWE
FIN
MYS
NLZ
BOL
HUN
SGP
URY
AUS
AUT
BEL
BRB
DEU
NLD
PRT
Paper
Textiles
CHL
COL
CRI
CUB
DOM
ECU
GTM
HND
JAM
NIC
PAN
SLV
VEN
ARG
BRA
CAN
HKG
IDN
KOR
MYS
PER
PHL
URY
AUS
AUT
BEL
BHS
BRB
CAN
DEU
ESP
FIN
FRA
GBR
GRC
HUN
ITA
JPN
NLD
NZL
PRI
PRT
SAU
SWE
TWN
USA
ARE
BLZ
BOL
SGP
THA
TTO
CHE
HTI
ZAF
CHL
COL
CRI
CUB
DOM
ECU
GTM
HND
NIC
PAN
PER
PRI
SLV
VEN
BZL
BOL
BRB
HTI
JAM
TTO
CHN
HKG
HUN
THA
TWN
BRA
CAN
CHE
DEU
IDN
KOR
PHL
SWE
USA
ARE
BEL
BHS
ESP
FRA
GRC
ITA
NLD
PRT
AUS
FIN
GBR
JPN
MYS
SGP
ZAF
ARG
AUT
NZL
SAU
URY
33
Table A3. Adjusted Rand indexes.
Agriculture
Chemicals
Machinery&Transport
Metals
Non-metallic minerals
Paper
Textiles
Agriculture
Chemicals
1.00
0.31
0.34
0.37
0.26
0.23
0.26
1.00
0.32
0.31
0.24
0.16
0.25
Machinery &
Transport
equipment
1.00
0.38
0.24
0.28
0.41
Metals
1.00
0.32
0.18
0.35
Non-metallic
minerals
1.00
0.17
0.23
Paper
1.00
0.25
Textiles
1.00
Industry partitions
vs. rest of industries
partitions
0.36
0.14
0.38
0.25
0.31
0.31
0.22
34
Table A4. Community detection process in the network of Mexican exporters' destinations using the 2000-2002 sample
ARG, BOL, CHL, COL, ECU, PER, PRY, URY, VEN
BLZ, CRI, CUB, DOM, GTM, HND, NIC, PAN, PRI, SLV
DMA, HTI, JAM, SUR, TTO
ABW, ANT, ATG, BHS, BMU, BRB, CYM, LCA, VCT, GUY
BEL, BRA, CAN, CHE, DEU, ESP, FRA, GBR, ITA, JPN, NLD, USA
ARE, EGY, JOR, KWT, LBN, SAU, SYR
DZA, LKA, MAR, PAK, TUR
DNK, FIN, NOR, SWE
AUT, CZE, HUN, IRL
NGA, VGB
CHN, HKG, IDN, IND, KOR, MYS, PHL, PRK, SGP, THA, TWN, VNM
AUS, NZL, ZAF
35